Automating the audit of the Brazilian electronic ballot operation: A new dataset for 6DoF pose estimation of the voter terminal based on domain randomization

Jefferson Medeiros Norberto, Kelvin Batista da Cunha, Marcondes Ricarte da Silva Júnior, Francisco Paulo Magalhaes Simões

Abstract: Performing detection and pose estimation of objects in six degrees of freedom (6-DoF) is a widely studied challenge in virtual and augmented reality, robotics and computer vision. For simulation and testing of the Brazilian voter terminal, its pose could allow automatic testing/auditing with robotics arms or virtual reality applications to simulate the voting process. For pose estimation using deep learning, it is necessary to generate large amounts of annotated real data, which is a costly task in time and resources. One way to avoid this issue is to create synthetic data through domain randomization, using 3D object modeling, to train the pose estimation technique with a reduced amount of annotated real data. In this work, domain randomization was utilized to generate a synthetic dataset, starting from a 3D model of the voter terminal, varying the lighting settings, camera position and distract insertion, to verify what impact this randomization has on training a single shot algorithm to perform the detection and pose estimation of this terminal in a different scenario. The new dataset with real and synthetic data from the voter terminal was built and will be publicly available.